New AI approach could improve railway fastener defect detection for smarter maintenance

Researchers have evaluated how Vision Transformers and convolutional neural networks can support faster and more accurate defect detection in railway track fasteners, a key task for smart railway maintenance. The study focuses on enhancing Non-Destructive Evaluation, or NDE, by using pre-trained deep learning models and transfer learning to identify irregularities without damaging transport infrastructure.

Predictive health management is increasingly important for rail networks because defects in track components can contribute to accidents, service disruptions, maintenance delays, and higher operating costs. Fasteners are small but essential parts of the rail system, helping secure the rail and maintain track integrity. If fastener defects are missed or detected too late, they can undermine the reliability of the larger infrastructure system. Traditional inspection workflows can be labor-intensive, and purely manual interpretation of images may be difficult to scale across large networks.

NDE and imaging techniques offer a promising route because they allow infrastructure to be inspected without destructive sampling. The challenge is that image-based defect detection still depends on models that can learn from limited and sometimes uneven datasets. In railway applications, labeled defect images may be scarce compared with more common computer vision domains. The new study addresses this constraint by using transfer learning, which allows models pre-trained on broader image data to be adapted for railway fastener defect detection.

The researchers assessed several machine learning models for enhancing NDE of railway track fasteners, including the Vision Transformer, or ViT, Data-efficient Image Transformer, or DeiT, VGG19, VGG16, and ResNet50. This comparison is important because convolutional neural networks, or CNNs, have long been widely used for image classification, while transformer-based vision models are increasingly being explored for tasks that require strong generalization and feature learning. According to the article, the application of transformers to railway fastener defect detection has remained relatively underexplored.

The results reported in the paper suggest that transformer-based models were the strongest performers. DeiT achieved 95.04% accuracy and ViT achieved 94.14%, both outperforming VGG16, which achieved 91.54%. The study also states that ViT and DeiT showed lower validation loss values, indicating stronger learning behavior and robustness in this task. These results highlight the potential value of transformer architectures for railway NDE, especially when paired with careful hyperparameter tuning.

The study does not suggest that CNNs have no role in this area. Instead, it presents a more nuanced picture of model choice. VGG models are described as a reliable alternative, while ResNet50 may be better suited for applications that prioritize computational efficiency over maximum accuracy. That distinction matters for real inspection systems, where model selection may depend on available computing resources, required response time, deployment platform, and the tolerance for missed or misclassified defects.

For railway operators and maintenance planners, the broader implication is that AI-assisted NDE could help move inspection practices toward more predictive and scalable maintenance. A model that can identify fastener defects from images with higher accuracy and stronger generalization could support earlier intervention, reduce reliance on manual screening, and help prioritize field inspections. In large rail systems, even incremental improvements in detection reliability may translate into better maintenance scheduling and fewer disruptions.

The findings also point toward future development of hybrid railway defect-detection models. The paper notes the promise of transformer-based approaches and suggests that larger datasets and future hybrid model designs could further improve performance. This is an important caveat, because real railway environments may include variations in lighting, camera angle, rail condition, fastener type, dirt, weather, and image quality that are difficult to capture fully in a limited dataset.

Further validation will still be needed before such methods can be considered ready for broad deployment across rail networks. Even so, the study offers a strong indication that transfer learning with ViT and DeiT models could strengthen AI-enhanced NDE for railway fastener defect detection. As transport infrastructure becomes more data-driven, tools that combine imaging, deep learning, and predictive maintenance may play a growing role in making railway systems safer, more reliable, and easier to maintain.

Reference
Author:
Samira Mohammadi ^a, Sasan Sattarpanah Karganroudi ^{b c}, Mehdi Adda ^a, Hussein Ibrahim ^c

Title of original paper:
Toward smart railway maintenance: AI-enhanced Non-Destructive Evaluation using Vision Transformers and CNNs for fastener defect detection

Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725000829

Journal:
Green Energy and Intelligent Transportation

DOI:
10.1016/j.geits.2025.100332

Affiliations:

^a Department of Mathematics, Computer Science and Engineering, Université du Québec à Rimouski, 1595 Bd Alphonse-Desjardins, Lévis, G6V 0A6, QC, Canada

^b Department of Mechanical Engineering, Université du Québec à Trois-Rivières, 575 Boul de l'Université, Drummondville, J2C 0R5, QC, Canada

^c Centre national intégré du manufacturier intelligent, Université du Québec à Trois-Rivières, 575 Boul de l'Université, Drummondville, J2C 0R5, QC, Canada